TL;DR: Google documents this directly: no specific optimisation is required for AI Overviews or AI Mode, and existing SEO fundamentals continue to apply. Eligibility follows ordinary indexing and snippet eligibility. The one mechanism worth understanding is query fan-out, where the system issues several related searches across subtopics, which means you compete on subtopics rather than on one query.
Key Takeaways
No special files, markup or AI-specific optimisation exist. Google states this in its own documentation for site owners. Anything claiming otherwise is inventing a requirement.
Eligibility is indexing plus snippet eligibility. A page that is not indexed, or not eligible to show a snippet, cannot appear as a supporting link.
Query fan-out changes what you target. The system may issue multiple related searches across subtopics, so coverage of the subtopics matters as much as the head query.
Preview controls are the real lever. nosnippet, max-snippet and data-nosnippet limit use as input, and they also limit your appearance in ordinary results.
AI Overviews and AI Mode differ. Google says they may use different models and techniques, so the links they show vary.
It is already in Search Console. Sites appearing in AI features are counted within overall Search performance data.
Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.
What Google says
This is the rare case in AI search where the operator publishes guidance for site owners. Read it before reading anything else, including this page.
Related reading: AI SEO vs Traditional SEO: Key Differences
The central statement
Google's documentation on AI features and your website says that while specific optimisation is not required for AI Overviews and AI Mode, all existing SEO fundamentals continue to be worthwhile.
That sentence does a lot of work. It rules out a category of advice, and it tells you where the effort goes: the things you were already doing.
What else it establishes
Point | Implication |
|---|---|
Both features may use query fan-out, issuing multiple related searches across subtopics | You compete across subtopics, not one query |
AI Mode and AI Overviews may use different models and techniques | Their link sets differ. Do not treat them as one surface |
Sites appearing in AI features are included in overall Search Console traffic | You already have data on this |
Preview controls limit use as direct input | The opt-out lever exists and has a cost |
Google also notes that with AI Overviews, people have been visiting a greater diversity of websites for more complex questions.
Why this page exists at all, then
Because the documentation is short and widely contradicted. A large amount of published advice recommends files, markup and rewrites that the operator says are unnecessary.
The useful contribution is to state what is documented, explain the one mechanism that genuinely changes your approach, and name what is being invented.
How to use this
Treat Google's page as primary and this one as a reading of it. Where they disagree, Google is right and this page needs updating.
That is an unusual thing for a vendor article to say and it is the correct relationship to a source like this.
Eligibility
The gate, and it is the thing to check before anything else.
The two conditions
A page that is not indexed cannot appear. A page that is not eligible to show a snippet cannot appear as a supporting link. Google states both directly.
That means AI Overview eligibility starts from ordinary Search eligibility, and a surprising number of pages fail at this gate for reasons nobody has looked at.
The checks
Check | How |
|---|---|
Is the page indexed? | URL Inspection in Search Console |
Does it show a snippet in ordinary results? | Search for a distinctive phrase from it |
Is a preview control suppressing the snippet? | Check for nosnippet, max-snippet, data-nosnippet |
Is it blocked in robots.txt? | Read the file, including on subdomains |
Is the content server-rendered? | View source and search for a body sentence |
The one people miss
Preview controls applied years ago for a different reason. A max-snippet limit set to protect content, or a nosnippet directive on a template, suppresses the snippet and therefore the eligibility.
Those directives usually predate AI features entirely and nobody revisits them. Worth a specific search across your templates.
What this means for prioritisation
If your pages are not indexed or not snippet-eligible, no amount of content restructuring matters. This is the equivalent of the crawler access check in the wider cluster: a binary gate that invalidates everything downstream.
The reassuring part
For most sites with functioning SEO, this gate is already passed. The check takes ten minutes and usually confirms you are fine, which is worth knowing before spending a quarter on content.
If it is not passed, you have found something more valuable than any optimisation advice.
Query fan-out
The one documented mechanism that genuinely changes how you think about targeting.
What Google describes
Both AI Overviews and AI Mode may use a query fan-out technique: issuing multiple related searches across subtopics and data sources to develop a response.
So the user asks one question, and the system runs several searches behind it.
Why that matters
You are not competing for one query. You are competing across whatever set of subtopic searches the system decides to run, and you may be pulled in on a subtopic you never considered a target.
A page that covers one narrow subtopic thoroughly can be retrieved for a broad question it would never have ranked for, and a page that ranks for the head term can be absent if the fan-out went somewhere it does not cover.
What follows for content
Implication | What to do |
|---|---|
Subtopic coverage matters | Build real depth across the sub-questions in your area, not one long page |
Narrow pages have a route in | A specific answer page can surface for broader questions |
Breadth without depth performs poorly | A page touching ten subtopics shallowly wins none of them |
The head term is not the target | Stop thinking in single-keyword terms entirely |
How to find your subtopics
Take a head question your buyers ask and write out every sub-question someone would need answered to act on it. Then check which of those you cover properly.
That list is usually longer than your content, and the gaps are plausible fan-out targets you currently cannot serve.
The connection to cluster thinking
This is a documented argument for topic clusters, which SEO practitioners have advocated for years on different reasoning. The mechanism is new; the practice is familiar.
If you already build genuine topic clusters with substantive pages per sub-question, you are well positioned, and that is worth saying internally rather than treating this as a new discipline.
The limit of what is known
Google says the technique may be used and does not publish how subtopics are selected. Do not trust anyone claiming to predict the fan-out for a given query; that is not knowable from outside.
Overviews and AI Mode
Frequently treated as one thing. Google says they may use different models and techniques, so the responses and links they show will vary.
The difference
AI Overviews | AI Mode | |
|---|---|---|
Where it appears | Within the ordinary results page, when Google's systems judge it adds value | A separate interface the user opts into |
How the user arrives | Automatically, on some queries | By choosing it |
Intent | A quick summary | Deeper exploration, comparison, follow-up |
Link sets | Differ from AI Mode's | Differ from Overviews' |
Why this matters practically
Appearing in one does not mean appearing in the other, so if you measure only the ordinary results page you may be missing the surface where considered comparison happens.
It also means a change in one is not evidence about the other.
Which to care about
Both, with different expectations. Overviews reach people who did not choose an AI experience; AI Mode reaches people who did and who are usually further into a question.
For considered purchases, AI Mode behaviour is closer to the assistant behaviour the rest of this cluster addresses.
The measurement complication
Google has said Search Console data does not include results from Search Labs experiments, so what you can see depends on what has moved out of experimental status.
Check Google's current documentation on what is counted before drawing conclusions from a Search Console trend, since this has changed over time and may change again.
The honest position
This is the part of the page most likely to date. The two surfaces are converging in some places and diverging in others, and the product keeps changing.
What holds is the underlying statement: they may use different models, so do not assume one tells you about the other.
The content work
Given that no AI-specific optimisation exists, what is left is familiar and still worth doing deliberately.
Practice | Why it helps here |
|---|---|
A direct answer near the top of the page | These features extract passages, not whole pages |
Headings that are the questions people ask | Matches the subtopic structure fan-out produces |
Self-contained paragraphs | A passage lifted out has to make sense alone |
Specifics rather than adjectives | A number can be quoted; an adjective cannot |
Genuine depth per subtopic | The fan-out argument |
Original data and first-hand experience | Google's own guidance emphasises helpful, reliable, original content |
Current pages, dated | Freshness matters more on fast-moving topics |
The one with the clearest mechanism
Passage extraction. If these systems pull specific passages rather than whole pages, then where your answer sits on the page determines whether it is available to be pulled.
An answer in the sixth paragraph, after context-setting, is less likely to be the extracted passage than the same answer in the first.
What this does not mean
Writing for machines. The same structural choices help a human scanning for an answer, which is why this advice predates AI features by years.
If a recommendation would make your page worse for a human reader, it is probably wrong.
Depth versus breadth
The fan-out mechanism argues for genuine depth across sub-questions rather than one comprehensive page. A long page covering ten subtopics thinly competes poorly on all ten.
That is an argument for more pages, each properly answering one question, which is also what the rest of this cluster recommends for assistants generally.
What to do first
Take your ten most commercially important pages and check whether a two-sentence answer could be lifted from the opening of each. Most sites fail that on most pages, and fixing it is a rewrite of the opening rather than the page.
Preview controls
The actual documented lever for limiting appearance, and a trap for anyone who used it before AI features existed.
What they do
Control | Effect |
|---|---|
| No text snippet shown for the page |
| Limits snippet length to n characters |
| Excludes a specific element on the page from snippets |
Google's documentation explains that these apply across its search surfaces, including AI Overviews and AI Mode, and will also limit how much of the content may be used as direct input.
The trade-off
These are the documented way to reduce use of your content in AI features. They also reduce your snippets in ordinary search results, because they are the same control.
There is no setting that keeps your ordinary snippets and removes you from AI features. That is the honest answer to a question many publishers ask.
Who this matters to
Situation | Consideration |
|---|---|
Publisher whose business is the click | A real trade-off worth modelling before acting |
Brand wanting to be recommended | You almost certainly want these off |
Site with historic controls nobody remembers | Check. They may be suppressing eligibility silently |
The audit worth running
Search your templates and CMS settings for all three directives. Find out when each was added and why.
The common finding is a max-snippet limit added years ago to protect content, still in place, now suppressing exactly the eligibility the team is trying to achieve.
The decision to make deliberately
If you are a brand wanting recommendations, you want full snippet eligibility. If you are a publisher whose revenue depends on clicks, the calculation genuinely differs and the data on both sides is thin.
Either way it should be a decision someone made recently, not a setting inherited from a different era.
What is not required
Google's documentation rules these out. They circulate anyway, including in material produced for this cluster.
Claimed requirement | Status |
|---|---|
An ai.json file | Not a standard. No provider documents reading it. Google says no new machine-readable files are needed |
An llms.txt file | A proposal, not a requirement. Published analysis found the overwhelming majority of such files received no requests at all |
AI-specific schema types | None exist. Standard structured data where relevant is the whole story |
An AI-friendly rewrite | Google says no specific optimisation is required |
Special meta tags for AI features | The preview controls are the only relevant directives, and they limit rather than enable |
LLM-friendly robots.txt as a distinct thing | robots.txt is robots.txt. What changes is which user agents you allow |
Why the ai.json claim keeps appearing
It sounds plausible. It is formatted like a real standard, it sits alongside robots.txt in the way sitemap.xml does, and the analogy to robots.txt makes it feel like a natural successor.
That analogy is the problem. robots.txt is honoured by crawlers because it is an established convention they implement. A file nobody implements is a file in your root directory doing nothing.
How to evaluate the next one
Ask who reads it. Not who proposed it, not who recommends it, but which named crawler implements it and where that is documented.
If there is no answer, the file does nothing. That test disposes of this entire category of recommendation.
The cost of implementing one anyway
Usually small and never zero: engineering time, a maintenance obligation, and a place where stale information can contradict your real pages.
At enterprise scale it is a sprint that could have gone to template fixes, which is the argument the enterprise article makes.
What to do with existing ones
If you already have an ai.json or llms.txt, leaving it is harmless. Just do not maintain it as if it matters, and do not let its existence substitute for the work that does.
Measuring it
This is the one AI surface where you already have first-party data, which makes it unlike everything else in this cluster.
What Search Console gives you
Google states that sites appearing in AI features are included in overall Search traffic in Search Console, the same as the rest of the results page.
So your impressions and clicks already include this. Google also publishes guidance on how AI features are counted and on analysing traffic changes, which is worth reading alongside your own data.
What this means practically
Advantage | Why it matters |
|---|---|
First-party | Not sampled, not modelled, not a vendor's estimate |
Already collected | No new measurement to set up |
Query-level | You can see which queries, not just an aggregate |
Free | No tooling decision required |
Compared to measuring assistants, where you sample prompts and infer a rate, this is direct observation. Use it.
The limitation
You cannot fully separate AI feature performance from ordinary results, because they are counted together. So you can see total performance and not a clean split.
That is a real constraint and it is still better data than any other surface offers. Read Google's current documentation on what is and is not counted before drawing conclusions.
What to watch
Signal | Reading |
|---|---|
Impressions steady, clicks falling on informational queries | Consistent with summary behaviour. Expected rather than a failure |
Both falling | A ranking or eligibility problem, not an AI one |
Clicks falling on transactional queries | Investigate. Less expected |
A specific page losing clicks while impressions hold | Check whether its snippet is being used in a summary |
The reporting implication
Judge informational content on impressions and position alongside clicks, rather than on clicks alone. The click rate on those queries is affected by something outside your control, and holding a content team to it is unfair.
That reframing is more useful than any optimisation on this page.
The traffic question
Worth addressing directly, because it is why most people arrive at this topic.
What the research shows
Pew Research Center analysed 68,879 searches and found that organic click-through roughly halved when an AI summary was present compared with searches without one, and that links inside the summary were clicked in around one percent of cases.
That is the clearest available measurement of the click cost, and it is substantial.
What Google says
That with AI Overviews, people have been visiting a greater diversity of websites for help with more complex questions.
These two statements are not necessarily contradictory. Fewer clicks overall and a wider spread of recipients can both be true, and they describe different things.
The honest synthesis
For informational queries where a summary appears, expect fewer clicks. For a site that was not previously winning those queries, a wider distribution of citations may represent an opportunity.
Whether you gain or lose depends on whether you were the incumbent beneficiary of those clicks, which is a different answer for a market-leading publisher than for a challenger.
What follows for strategy
If you were | Then |
|---|---|
The top result on informational queries | You are likely losing clicks. Diversify toward queries where summaries appear less |
Not ranking well on those queries | Citation in a summary is a route you did not previously have |
Dependent on informational traffic for revenue | This is a business model question, not an SEO one |
Selling considered purchases | The clicks were never the whole mechanism. Being cited still carries weight |
What not to conclude
That AI Overviews are simply bad for everyone. The evidence shows a click cost and a redistribution, and which side of that you land on depends on your starting position.
Reporting it as a uniform catastrophe is inaccurate and it also removes any route to a strategy.
Six mistakes
1. Implementing files nobody reads. Google states no new machine-readable files are needed. An ai.json sitting in your root directory is doing nothing, and the engineering time had an alternative use.
2. Not checking snippet eligibility. A page that cannot show a snippet cannot appear as a supporting link. Preview controls set years ago for other reasons are the usual cause, and nobody revisits them.
3. Treating AI Overviews and AI Mode as one surface. Google says they may use different models and techniques, so appearing in one tells you little about the other.
4. Optimising for the head query. Query fan-out means the system may issue several subtopic searches. Depth across sub-questions matters more than one page targeting the main term.
5. Judging content teams on click-through alone. Published research found organic click-through roughly halved when a summary was present. That is outside the team's control and holding them to it is unfair.
6. Ignoring Search Console. This is the one AI surface with first-party data, already collected, at query level, for free. Most teams measuring assistants painstakingly have never looked at it for this.
The pattern
Two are inventions, two are failures to look at available information, and two are misreadings of the mechanism. None requires new tooling to fix.
The one to do today
Search your templates for nosnippet, max-snippet and data-nosnippet. Find out when each was added and why.
If one of them is suppressing eligibility on pages you want cited, that is a single-line fix for a problem that no amount of content work would have solved.
FAQs
How do I optimise for Google AI Overviews?
Google states that no specific optimisation is required and that existing SEO fundamentals continue to apply. Ensure pages are indexed and snippet-eligible, answer questions directly near the top of the page, and build genuine depth across subtopics.
Do I need an ai.json or llms.txt file?
No. Google's documentation says no new machine-readable files are needed for its AI features, and no provider documents reading ai.json. Published analysis found the overwhelming majority of llms.txt files received no requests at all.
What makes a page eligible to appear?
Ordinary Search eligibility. A page that is not indexed, or not eligible to show a snippet, cannot appear as a supporting link. Check for preview controls that may be suppressing snippets, often added years earlier for unrelated reasons.
What is query fan-out?
Google describes it as issuing multiple related searches across subtopics and data sources to develop a response. It means you compete across subtopics rather than one query, which argues for genuine depth per sub-question rather than one broad page.
Are AI Overviews and AI Mode the same?
No. Overviews appear within ordinary results when Google judges they add value; AI Mode is a separate interface users opt into. Google says they may use different models and techniques, so the links they show vary.
Can I appear in normal results but not AI Overviews?
Not selectively. The preview controls that limit use in AI features are the same ones that control your ordinary snippets, so reducing one reduces the other. There is no setting that separates them.
How do I measure AI Overview performance?
Search Console includes sites appearing in AI features within overall Search traffic. You cannot fully separate it from ordinary results, and it is still first-party query-level data, which no other AI surface offers.
Will AI Overviews reduce my traffic?
Research by Pew found organic click-through roughly halved when a summary was present, with in-summary links clicked around one percent of the time. Whether you lose depends on whether you were previously winning those clicks.
Where to start
Search your templates for nosnippet, max-snippet and data-nosnippet. If one of them is suppressing snippets on pages you want cited, that is a one-line fix for a problem no content work would solve.
Then open Search Console. You already have first-party query-level data on this surface, which is more than any other part of AI search offers, and most teams have never looked at it for this purpose.
The order
Check eligibility. Indexed, snippet-eligible, server-rendered, not blocked.
Audit preview controls. Find out what was set, when and why.
Move answers up. The opening hundred words of your ten most important pages.
Map subtopics. For each head question, list the sub-questions and check your coverage.
Read Google's documentation directly. It is short and it is the primary source.
And do not build ai.json, llms.txt, or anything else nobody reads.
If subtopic coverage is the gap
Query fan-out rewards genuine depth across sub-questions, which means more pages each answering one thing properly. Producing those consistently is where teams stall, and Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.
The eligibility and preview control work is engineering, and it comes first.
Summary
Google documents this directly: no specific optimisation is required for AI Overviews or AI Mode, and ordinary SEO fundamentals still apply. Eligibility follows ordinary indexing and snippet eligibility, so a page that cannot show a snippet cannot appear, and preview controls set years ago are a common silent cause. The mechanism worth understanding is query fan-out, where the system issues several related searches across subtopics, which argues for real depth across sub-questions rather than one page targeting a head term. Overviews and AI Mode may use different models, so do not treat them as one surface. The files and markup widely recommended for this are not read by anyone, and the test for the next proposal is simple: which named crawler implements it, and where is that documented. Finally, this is the one AI surface where Search Console already gives you first-party query-level data, which makes it the cheapest thing in the entire field to measure.